Abstract: Techniques for determining when speech is directed at another individual of a dialog, and storing a representation of such user-directed speech for use as context when processing subsequently-received system-directed speech are described. A system receives audio data and/or video data and determines therefrom that speech in the audio data is user-directed. Based on this, the system determine whether the speech is able to be used to perform an action by the system. If the speech is able to be used to perform an action, the system stores a natural language representation of the speech. Thereafter, when the system receives system-directed speech, the system generates a rewrite of a natural language representation of the system-directed speech based on the previously-received user-directed speech. The system then determines output data responsive to the system-directed speech using the rewritten natural language representation.

Abstract: A method of decoding an original speech signal for hybrid adversarial-parametric speech synthesis comprising: (a) receiving quantized original linear prediction coding parameters estimated by applying linear prediction coding analysis filtering to an original speech signal and a quantized compressed representation of a residual of the original speech signal; (b) dequantizing the original linear prediction coding parameters and the compressed representation of the residual; (c) inputting the dequantized compressed representation of the residual into a decoder part of a Generator for applying adversarial mapping from the compressed residual domain to a fake (first) signal domain; (d) outputting, by the decoder part of the Generator, a fake speech signal; (e) applying linear prediction coding analysis filtering to the fake speech signal for obtaining a corresponding fake residual; (f) reconstructing the original speech signal by applying linear prediction coding cross-synthesis filtering to the fake residual and

Abstract: Apparatuses, methods and storage medium for computing including determination of work placement on processor cores are disclosed herein. In embodiments, an apparatus may include one or more processors, devices, and/or circuitry to identify a favored core of the processor cores. The one or more processors, devices, and/or circuitry may be configured to determine whether to migrate a thread to or from the favored core. In some embodiments, the determination may be by a process executed by a driver and/or by an algorithm executed by a power control unit of the processor.

Abstract: A method of decoding an original speech signal for hybrid adversarial-parametric speech synthesis comprising: (a) receiving quantized original linear prediction coding parameters estimated by applying linear prediction coding analysis filtering to an original speech signal and a quantized compressed representation of a residual of the original speech signal; (b) dequantizing the original linear prediction coding parameters and the compressed representation of the residual; (c) inputting the dequantized compressed representation of the residual into a decoder part of a Generator for applying adversarial mapping from the compressed residual domain to a fake (first) signal domain; (d) outputting, by the decoder part of the Generator, a fake speech signal; (e) applying linear prediction coding analysis filtering to the fake speech signal for obtaining a corresponding fake residual; (f) reconstructing the original speech signal by applying linear prediction coding cross-synthesis filtering to the fake residual and

Abstract: Embodiments are directed to a companding method and system for reducing coding noise in an audio codec. A compression process reduces an original dynamic range of an initial audio signal through a compression process that divides the initial audio signal into a plurality of segments using a defined window shape, calculates a wideband gain in the frequency domain using a non-energy based average of frequency domain samples of the initial audio signal, and applies individual gain values to amplify segments of relatively low intensity and attenuate segments of relatively high intensity. The compressed audio signal is then expanded back to the substantially the original dynamic range that applies inverse gain values to amplify segments of relatively high intensity and attenuating segments of relatively low intensity. A QMF filterbank is used to analyze the initial audio signal to obtain a frequency domain representation.

Abstract: The present disclosure relates to methods for processing a decoded audio signal and for selectively applying speech/dialog enhancement to the decoded audio signal. The present disclosure also relates to a method of operating a headset for computer-mediated reality. A method of processing a decoded audio signal comprises obtaining a measure of a cognitive load of a listener that listens to a rendering of the audio signal, determining whether speech/dialog enhancement shall be applied based on the obtained measure of the cognitive load, and performing speech/dialog enhancement based on the determination. A method of operating a headset for computer-mediated reality comprises obtaining eye-tracking data of a wearer of the headset, determining a measure of a cognitive load of the wearer of the headset based on the eye-tracking data, and outputting an indication of the cognitive load of the wearer of the headset.

Abstract: Techniques for generating a prompt for a language model to determine an action responsive to a user input, are described. In some embodiments, the system receives a user input, determines one or more application programming interfaces (APIs) configured to perform actions that are relevant to the user input and exemplars representing examples of using the APIs with respect to user inputs similar to the current user input. The system further determines device states of devices that are determined to be related to the user input and also determines other contextual information (e.g., weather information, time of day, geographic location, etc.). The system generates a prompt including the user input, the APIs, the exemplars, the device states, and the other contextual information. A language model processes the prompt to determine an action responsive to the user input and the system causes performance of the action.

Abstract: Embodiments are directed to a companding method and system for reducing coding noise in an audio codec. A compression process reduces an original dynamic range of an initial audio signal through a compression process that divides the initial audio signal into a plurality of segments using a defined window shape, calculates a wideband gain in the frequency domain using a non-energy based average of frequency domain samples of the initial audio signal, and applies individual gain values to amplify segments of relatively low intensity and attenuate segments of relatively high intensity. The compressed audio signal is then expanded back to the substantially the original dynamic range that applies inverse gain values to amplify segments of relatively high intensity and attenuating segments of relatively low intensity. A QMF filterbank is used to analyze the initial audio signal to obtain a frequency domain representation.

Abstract: Systems and methods are disclosed to implement a neural network training system to train a multitask neural network (MNN) to generate a low-dimensional entity representation based on a sequence of events associated with the entity. In embodiments, an encoder is combined with a group of decoders to form a MNN to perform different machine learning tasks on entities. During training, the encoder takes a sequence of events in and generates a low-dimensional representation of the entity. The decoders then take the representation and perform different tasks to predict various attributes of the entity. As the MNN is trained to perform the different tasks, the encoder is also trained to generate entity representations that capture different attribute signals of the entities. The trained encoder may then be used to generate semantically meaningful entity representations for use with other machine learning systems.

Abstract: Apparatuses, methods and storage medium for computing including determination of work placement on processor cores are disclosed herein. In embodiments, an apparatus may include one or more processors, devices, and/or circuitry to identify a favored core of the processor cores. The one or more processors, devices, and/or circuitry may be configured to determine whether to migrate a thread to or from the favored core. In some embodiments, the determination may be by a process executed by a driver and/or by an algorithm executed by a power control unit of the processor.

Abstract: Systems and methods are disclosed to implement a neural network training system to train a multitask neural network (MNN) to generate a low-dimensional entity representation based on a sequence of events associated with the entity. In embodiments, an encoder is combined with a group of decoders to form a MNN to perform different machine learning tasks on entities. During training, the encoder takes a sequence of events in and generates a low-dimensional representation of the entity. The decoders then take the representation and perform different tasks to predict various attributes of the entity. As the MNN is trained to perform the different tasks, the encoder is also trained to generate entity representations that capture different attribute signals of the entities. The trained encoder may then be used to generate semantically meaningful entity representations for use with other machine learning systems.

Abstract: Described herein is a method of low-bitrate coding of audio data and generating enhancement metadata for controlling audio enhancement of the low-bitrate coded audio data at a decoder side, including the steps of: (a) core encoding original audio data at a low bitrate to obtain encoded audio data; (b) generating enhancement metadata to be used for controlling a type and/or amount of audio enhancement at the decoder side after core decoding the encoded audio data; and (c) outputting the encoded audio data and the enhancement metadata. Described is further an encoder configured to perform said method. Described is moreover a method for generating enhanced audio data from low-bitrate coded audio data based on enhancement metadata and a decoder configured to perform said method.

Abstract: Apparatuses, methods and storage medium for computing including determination of work placement on processor cores are disclosed herein. In embodiments, an apparatus may include one or more processors, devices, and/or circuitry to identify a favored core of the processor cores. The one or more processors, devices, and/or circuitry may be configured to determine whether to migrate a thread to or from the favored core. In some embodiments, the determination may be by a process executed by a driver and/or by an algorithm executed by a power control unit of the processor.

Abstract: Embodiments are directed to a companding method and system for reducing coding noise in an audio codec. A method of processing an audio signal includes the following operations. A system receives an audio signal. The system determines that a first frame of the audio signal includes a sparse transient signal. The system determines that a second frame of the audio signal includes a dense transient signal. The system compresses/expands (compands) the audio signal using a companding rule that applies a first companding exponent to the first frame of the audio signal and applies a second companding exponent to the second frame of the audio signal, each companding exponent being used to derive a respective degree of dynamic range compression and expansion for a corresponding frame. The system then provides the companded audio signal to a downstream device.

Abstract: Respective initial feature sets are obtained for the nodes of a graph in which the nodes represent instances of entity types and edges represent relationships. Using the initial feature sets and the graph, a graph convolutional model is trained to generate one or more types of predictions. In the model, a representation of a particular node at a particular hidden layer is based on aggregated representations of neighbor nodes, and an embedding produced at a final hidden layer is used as input to a prediction layer. The trained model is stored.

Abstract: Described herein is a method of generating, in a dynamic range reduced domain, an enhanced multi-channel audio signal from an audio bitstream including a multi-channel audio signal, wherein the multi-channel audio signal comprises two or more channels, and wherein the method includes jointly enhancing the two or more channels of the dynamic range reduced raw multi-channel audio signal using a multi-channel Generator of a Generative Adversarial Network setting. Described herein are further a method for training a multi-channel Generator in a dynamic range reduced domain in a Generative Adversarial Network setting, an apparatus for generating, in a dynamic range reduced domain, an enhanced multi-channel audio signal from an audio bitstream including a multi-channel audio signal, respective systems and a computer program product.

Abstract: Respective initial feature sets are obtained for the nodes of a graph in which the nodes represent instances of entity types and edges represent relationships. Using the initial feature sets and the graph, a graph convolutional model is trained to generate one or more types of predictions. In the model, a representation of a particular node at a particular hidden layer is based on aggregated representations of neighbor nodes, and an embedding produced at a final hidden layer is used as input to a prediction layer. The trained model is stored.

Abstract: Respective initial feature sets are obtained for the nodes of a graph in which the nodes represent instances of entity types and edges represent relationships. Using the initial feature sets and the graph, a graph convolutional model is trained to generate one or more types of predictions. In the model, a representation of a particular node at a particular hidden layer is based on aggregated representations of neighbor nodes, and an embedding produced at a final hidden layer is used as input to a prediction layer. The trained model is stored.

Abstract: Embodiments are directed to a companding method and system for reducing coding noise in an audio codec. A compression process reduces an original dynamic range of an initial audio signal through a compression process that divides the initial audio signal into a plurality of segments using a defined window shape, calculates a wideband gain in the frequency domain using a non-energy based average of frequency domain samples of the initial audio signal, and applies individual gain values to amplify segments of relatively low intensity and attenuate segments of relatively high intensity. The compressed audio signal is then expanded back to the substantially the original dynamic range that applies inverse gain values to amplify segments of relatively high intensity and attenuating segments of relatively low intensity. A QMF filterbank is used to analyze the initial audio signal to obtain a frequency domain representation.

Abstract: Described herein is a method of waveform decoding, the method including the steps of: (a) receiving, by a waveform decoder, a bitstream including a finite bitrate representation of a source signal; (b) waveform decoding the finite bitrate representation of the source signal to obtain a waveform approximation of the source signal; (c) providing the waveform approximation of the source signal to a generative model that implements a probability density function, to obtain a probability distribution for a reconstructed signal of the source signal; and (d) generating the reconstructed signal of the source signal based on the probability distribution. Described are further a method and system for waveform coding and a method of training a generative model.